A Hybrid Machine Learning Approach for Synthetic Data Generation with Post Hoc Calibration for Clinical Tabular Datasets

TL;DR

This work tackles data scarcity and privacy constraints in healthcare by introducing a hybrid synthetic data generation framework that fuses five augmentation techniques—noise injection, interpolation, GMM sampling, CVAE sampling, and SMOTE—with a reinforcement‑learning–driven weighting mechanism and a multi‑stage calibration suite (moment matching, full histogram, soft/adaptive soft, iterative soft histogram). The approach achieves near‑real distributional fidelity and strong downstream utility across oncology and cardiology datasets, with Wasserstein distances as low as $W(P,Q) \approx 0.001$ and KS around $0.01$, NNAA near 50%, and classification performance up to $\approx 94\%$ accuracy and $F1 \approx 93\%$. Benchmarking against SDV models shows competitive or superior marginal fidelity, while balancing joint distribution preservation via calibration, enabling privacy‑preserving AI in sensitive healthcare settings. These results demonstrate the framework’s potential to accelerate safe AI development in data‑constrained environments and motivate future work on integrating diffusion models and standardized evaluation protocols for broader, multi‑modal health data.

Abstract

Healthcare research and development face significant obstacles due to data scarcity and stringent privacy regulations, such as HIPAA and the GDPR, restricting access to essential real-world medical data. These limitations impede innovation, delay robust AI model creation, and hinder advancements in patient-centered care. Synthetic data generation offers a transformative solution by producing artificial datasets that emulate real data statistics while safeguarding patient privacy. We introduce a novel hybrid framework for high-fidelity healthcare data synthesis integrating five augmentation methods: noise injection, interpolation, Gaussian Mixture Model (GMM) sampling, Conditional Variational Autoencoder (CVAE) sampling, and SMOTE, combined via a reinforcement learning-based dynamic weight selection mechanism. Its key innovations include advanced calibration techniques -- moment matching, full histogram matching, soft and adaptive soft histogram matching, and iterative refinement -- that align marginal distributions and preserve joint feature dependencies. Evaluated on the Breast Cancer Wisconsin (UCI Repository) and Khulna Medical College cardiology datasets, our calibrated hybrid achieves Wasserstein distances as low as 0.001 and Kolmogorov-Smirnov statistics around 0.01, demonstrating near-zero marginal discrepancy. Pairwise trend scores surpass 90%, and Nearest Neighbor Adversarial Accuracy approaches 50%, confirming robust privacy protection. Downstream classifiers trained on synthetic data achieve up to 94% accuracy and F1 scores above 93%, comparable to models trained on real data. This scalable, privacy-preserving approach matches state-of-the-art methods, sets new benchmarks for joint-distribution fidelity in healthcare, and supports sensitive AI applications.

Figures (10)

• Figure 1: Workflow of the hybrid synthetic data generation framework. The figure illustrates data preprocessing, hybrid generation via five techniques, sequential application of the five calibration methods, and evaluation stages. The five post-calibration methods are depicted as iterative refinements to align synthetic distributions with original data.
• Figure 2: Projection of real vs. synthetic Breast Cancer (Original) data in PCA, t-SNE, and UMAP spaces. Real data points and synthetic data points are plotted together to visualize overlap; ideally, synthetic points align with the real data distribution in each projection.
• Figure 3: Correlation matrices of features for the Breast Cancer (Original) dataset: (left) real data, (right) synthetic data (post-calibration). Color intensity indicates the strength of Pearson correlations between feature pairs.
• Figure 4: Feature density distributions for selected features in the Breast Cancer (Original) dataset, comparing real (solid line) vs. synthetic (dashed line) data. Each subplot represents one feature's distribution.
• Figure 5: Real vs. synthetic data projections for the Breast Cancer (Diagnostic) dataset. The synthetic data shown here uses the calibrated hybrid model output.